AI math gains raise security and research concerns

OpenAI and Prof. Amnon Shashua’s doubleAI have reported progress on mathematical problems long considered out of reach, prompting excitement and unease across the scientific community. doubleAI said it solved the longest path problem in graph theory, which had remained unsolved for 38 years, using AI agents at a cost of less than ?,000. Shashua said the company introduced an algebraic formula that raises a mathematical constant from 4/3 to 1.6, with potential implications for error correction, quantum computing and chip design.

OpenAI separately released a database of 722 studies with proofs for open problems in mathematics and computer science, following its reported work on the Navier-Stokes problem. Some proofs have been verified through Lean. The company said many results came from one AI agent and often a single prompt, a sharp contrast with the Navier-Stokes effort, which used 10,000 agents.

The advances have sharpened concerns about research norms and security. 25 winners of the Fields Medal warned that using open problems as benchmarks could encourage mass production of answers without enough time to understand, connect or credit the underlying ideas. Timothy Gowers argued that the larger risk may be a shrinking human research community, while others warned that stronger mathematical AI could affect encryption, communications and cyber capabilities.